Log in
|
Resident of the world, traveling the road of life
|
When you’ve read this post, go watch the worked examples on YouTube. [YouTube]
Our good friend Aron Peterson, a.k.a. Shokunin Studio, is a media production professional. He experiments with AI image and video generators. But he also knows what professional quality is.
Google Veo 3 is the new hotness in video generation. It does sound as well as video. But can it follow a script? You can’t do professional work if you can’t follow a script.
AI shills keep claiming that video generators like Veo 3 can do professional work. Aron thought: let’s take the AI pumpers at their word. Let’s use Veo like an AV professional would do.
As we’ve seen over the past two weeks, Veo will not take direction. It has no consistency. It keeps making stuff up.
When AI promoters tell you that video generators will replace human actors any moment, they are lying.
Veo is really bad at keeping track of who says what in a scene. It gets confused and it drops to silence, it puts in nonsense text subtitles, or it drops in a random scene you didn’t ask for.
Veo can’t do voiceovers — it keeps making the actor speak the lines. Veo randomly switches between voiceover, physical monologue, subtitles with no voiceover, or complete silence.
Go to the video and listen to the audio quality in the Day 15 clip that remakes a scene from Godfather 2 (3:34 on). The Veo-generated sound is like it’s down a tunnel — like really bad compression artefacts or noise reduction turned up too high. You can hear odd effects in a lot of Veo audio, but it’s really obvious in this segment. Veo audio also clips a lot — particularly on sound effects.
Veo just can’t act. Aron tried doing a gangster movie, Sexy Beast by Jonathan Glazer:
Veo’s version of Don Logan wasn’t intimidating or a sociopath. It came across as a newsreader on British TV. There’s no tension. Generated actors are always bland no matter how well they generate.
For still image generators, it took a couple of years before they could render text that wasn’t random letter-shaped gibberish — and even then they still fail. The Veo developers didn’t bother doing that work, and Veo makes a complete mess of putting text on a screen.
Veo does plausible output — but only if you don’t really care what it spits out. If you’ll take any old garbage that looks plausible for seven seconds, Veo’s just the thing!
A lot of AI fans are still telling Aron it can’t be that stupid, he must be prompting it wrong. But they still haven’t posted a prompt that will work reliably for other people. Perhaps you can respond to this post, and show us what a good prompt engineer you are.
Now go watch the video for the fun stuff. And check out last week’s Veo fails and the week before.